Google puts Gemini managed agents on the free tier and adds hard token caps

Google expanded Managed Agents in the Gemini API on Tuesday, July 28, and buried the biggest pricing move in the middle of a feature list: managed agents are now available on free-tier Gemini API projects, with no active billing required. A developer can now stand up an autonomous, sandboxed agent workload without entering a credit card. The trial cost of agentic compute on Google's platform is, for the moment, zero.

The free tier changes the trial math

Until now, kicking the tires on a managed agent meant a billed project, which meant procurement friction for exactly the developers Google wants experimenting. Removing that gate is a classic acquisition play, and it extends a pricing offensive Google has been running all month. The company cut Gemini 3.6 Flash prices on July 21 and positioned Flash-Lite as a budget agent model the same week. Free managed agents are the logical next step: get the workload running first, meter it later.

Alongside the tier change, Gemini 3.6 Flash becomes the default model for Managed Agents. Projects using the agent alias antigravity-preview-05-2026 get the new model with no code change, and 3.5 Flash and 3.5 Flash-Lite remain selectable through agent_config.model, per Google's developer blog post.

Caps that pause instead of bills that surprise

The operational additions read like answers to the most common complaints about running agents in production. A new max_total_tokens parameter caps total consumption across input, output and thinking tokens. When an agent hits the cap, execution pauses safely and returns a status of 'incomplete', with state preserved; the run can be resumed through previous_interaction_id. That turns runaway-agent cost from an incident into a checkpoint, and it makes spend governance a first-class platform feature rather than a wrapper someone builds on top. Anyone who has watched an agent burn through a budget in a retry loop will recognize what problem this is solving, and why finance teams have been the quiet blockers on agent rollouts.

Environment hooks go in the same direction. A .agents/hooks.json file supports pre_tool_execution and post_tool_execution hooks that can block, lint or audit tool calls inside the sandbox. Scheduled execution rounds out the release, binding agents, environments, prompts and cron expressions so recurring autonomous tasks run without a human trigger. Sandboxes persist between runs with a seven-day default TTL, a figure that so far appears only in Google's own post. Coverage at Crypto Briefing relayed the expansion but added no independent detail on retention.

What Google is really selling

Taken together, the release is less about any single feature than about where Google wants the agent abstraction to live. Managed Agents handle the loop, the sandbox, the scheduling and now the cost ceiling. Developers bring prompts and tools. That is a direct bid against the do-it-yourself orchestration stack, and the free tier is the funnel into it.

For the agentic AI as a service market, this sets a new floor. When the largest API providers hand out managed agent runtime for free at trial scale, standalone platforms selling the same plumbing have to justify their existence on governance depth, cross-model flexibility or enterprise controls rather than on the runtime itself. The caps-and-hooks combination is also a preview of where buyer expectations are heading: pause-and-resume budget enforcement and auditable tool calls will stop being differentiators and start being table stakes. Google gave away the on-ramp on Tuesday. The meter, as always, starts once the workload is too embedded to move.